STAT 301 Final Project

Predicting Placement Status in Students

Group 35:

• Vedant Kalyani
• Timothy Gao
• Ryan Lowe
• Aryan Jain

Course: STAT 301

Date: 16th April 2025

1. Introduction

Securing a Job Placement after Graduation is a key concern for us students in today's compepetive job market. Many companies and recruiters assess candidates on a variety of criteria beyond just academic performance, such as soft skills, extracurricular involvement, and technical certifications. With access to data that captures a range of these attributes, it becomes valuable to explore whether we can predict placement outcomes for students based on these attributes.

Research Question:

How can academic performance, soft skills, internships, and related personal attributes be leveraged to predict a student’s probability of receiving a job placement upon graduation?

We chose this research question because it allows us to build a model that predicts the likelihood of a student being placed after graduation using relevant academic and personal factors. This approach allows us to apply predictive modeling techniques while ensuring the model can generalize well. As students who will soon be looking for placements ourselves, it will help provide students like us with more insights into improving career readiness and ensure the best chance at a placement!

2. Data Description

The table below provides a summary of the dataset, including variable names, data types, and descriptions.

Variable	Data Type	Description
StudentID	integer	Student ID
CGPA	float	Overall grades achieved by the student
Internships	integer	Number of internships a student has done
Projects	integer	Number of projects a student has done
Workshops.Certifications	integer	Number of online certifications/workshops completed
AptitudeTestScore	integer	Aptitude test score to assess logical thinking
SoftSkillsRating	numeric	Communication rating affecting placement chances
ExtracurricularActivities	boolean	Whether the student participated in extracurriculars
PlacementTraining	boolean	Whether the student received placement training
SSC_Marks	integer	Senior Secondary Marks
HSC_Marks	integer	Higher Secondary Marks
PlacementStatus	boolean	Whether a student was placed (TRUE/FALSE)

Dataset overview: The file comprises 10,000 observations across 12 variables. Predictor variables span academic metrics (CGPA, SSC_Marks, HSC_Marks), experiential factors (Internships, Projects, Workshops_Certifications), and personal development indicators (SoftSkillsRating, ExtracurricularActivities, PlacementTraining). The binary target variable, PlacementStatus, indicates whether each student secured a placement.

Source: Kaggle — “Placement Prediction Dataset” (Ruchika Kumbhar) https://www.kaggle.com/datasets/ruchikakumbhar/placement-prediction-dataset/data.

3. EDA & Visualization

#Load libraries
library(tidyverse)
library(broom)
library(pROC)

── Attaching core tidyverse packages ──────────────────────── tidyverse 2.0.0 ──
✔ dplyr     1.1.4     ✔ readr     2.1.5
✔ forcats   1.0.0     ✔ stringr   1.5.1
✔ ggplot2   3.5.1     ✔ tibble    3.2.1
✔ lubridate 1.9.3     ✔ tidyr     1.3.1
✔ purrr     1.0.2     
── Conflicts ────────────────────────────────────────── tidyverse_conflicts() ──
✖ dplyr::filter() masks stats::filter()
✖ dplyr::lag()    masks stats::lag()
ℹ Use the conflicted package (<http://conflicted.r-lib.org/>) to force all conflicts to become errors
Type 'citation("pROC")' for a citation.


Attaching package: ‘pROC’


The following objects are masked from ‘package:stats’:

    cov, smooth, var

# 3  Data Loading & Initial Cleaning
#load the data via a github link
placement_df <- read.csv("https://raw.githubusercontent.com/Vedkal19/placement-data/main/placementdata.csv")

# Convert character columns to logical binary 1/0 for inference
placement_df <- placement_df %>%
  mutate(
    PlacementStatus = ifelse(PlacementStatus == "Placed", 1, 0),
    PlacementTraining = ifelse(PlacementTraining == "Yes", 1, 0),
    ExtracurricularActivities = ifelse(ExtracurricularActivities == "Yes", 1, 0)
  )

#Checking the structure of the dataset
str(placement_df)

#Checking for missing values
colSums(is.na(placement_df))

'data.frame':	10000 obs. of  12 variables:
 $ StudentID                : int  1 2 3 4 5 6 7 8 9 10 ...
 $ CGPA                     : num  7.5 8.9 7.3 7.5 8.3 7 7.7 7.7 6.5 7.8 ...
 $ Internships              : int  1 0 1 1 1 0 1 2 1 1 ...
 $ Projects                 : int  1 3 2 1 2 2 1 1 1 3 ...
 $ Workshops.Certifications : int  1 2 2 2 2 2 1 0 0 2 ...
 $ AptitudeTestScore        : int  65 90 82 85 86 71 76 85 84 79 ...
 $ SoftSkillsRating         : num  4.4 4 4.8 4.4 4.5 4.2 4 3.5 3.9 4.4 ...
 $ ExtracurricularActivities: num  0 1 1 1 1 1 0 1 0 1 ...
 $ PlacementTraining        : num  0 1 0 1 1 0 0 1 1 1 ...
 $ SSC_Marks                : int  61 78 79 81 74 55 62 59 75 85 ...
 $ HSC_Marks                : int  79 82 80 80 88 66 65 72 71 86 ...
 $ PlacementStatus          : num  0 1 0 1 1 0 0 0 0 1 ...

StudentID

0

CGPA

0

Internships

0

Projects

0

Workshops.Certifications

0

AptitudeTestScore

0

SoftSkillsRating

0

ExtracurricularActivities

0

PlacementTraining

0

SSC_Marks

0

HSC_Marks

0

PlacementStatus

0

The dataset is already tidy:

• Each row represents one student/ observation
• Each column represents one variable
• Each cell contains a single value
• There are no missing values

# 3.1  Soft‑Skills vs Aptitude Test Score
ggplot(placement_df, aes(x = SoftSkillsRating, y = AptitudeTestScore, color = PlacementStatus)) +
  geom_point() +
  labs(title = "Soft Skills Rating vs. Aptitude Score",
       x = "Soft Skills Rating",
       y = "Aptitude Test Score",
       color = "Placement Status")

Interpretation: The scatter plot above helps us explore the relationship between Soft Skills Rating, Aptitude Test Score, and Placement Status. This helps us investigate two key factors in relation to Placement Status. We can see that the students with a higher Aptitude Test score and Soft Skills rating seem to be placed more compared to students that have lower rating/ score in either feature as there is a higher density of blue dots in the top right of the graph compared to the bottom left. This helps us understand that these two factors might be quite important in predicting whether a student gets placed or not.

# 3.2  Box‑Plots: Numeric Features by Placement Status
# Reshaping numeric explanatory variables for plotting
numeric_vars <- placement_df %>%
  select(CGPA, SoftSkillsRating, AptitudeTestScore, SSC_Marks, HSC_Marks, PlacementStatus) %>%
  pivot_longer(cols = -PlacementStatus, names_to = "Variable", values_to = "Value")

ggplot(numeric_vars, aes(x = factor(PlacementStatus), y = Value, fill = factor(PlacementStatus))) +
  geom_boxplot(alpha = 1.5, width = 0.8) +
  facet_wrap(~ Variable, scales = "free", ncol = 3) +
  labs(
    title = "Boxplots of Numeric Features by Placement Status",
    x = "Placement Status (0 = Not Placed, 1 = Placed)",
    y = "Value"
  ) +
  theme_minimal()

Interpretation: The boxplots above help us explore how the numeric features in the data differ for students who were placed vs students who were not placed. We can see that for HSC Marks, and Aptitude Test Score, students who had higher values have a much higher placement rate than students who have lower values. While SSC Marks, CGPA, and Soft Skills Rating don't have the same level of discreptency between high value vs low value scores, they also do show that students who scored higher in these metrics did have a slightly higher placement rate than students who scored low in these metrics. These visual differences suggest that these variables will probably be important predictors in determining Placement Status, which can guide our feature selection for building our predictive model.

# 3.3  Correlation Heat‑Map
numeric_vars <- placement_df %>%
  select(CGPA, SSC_Marks, HSC_Marks, AptitudeTestScore, SoftSkillsRating, Projects)

cor_matrix <- round(cor(numeric_vars), 2)

# Convert to long format for ggplot
cor_long <- as.data.frame(as.table(cor_matrix))

ggplot(cor_long, aes(Var1, Var2, fill = Freq)) +
  geom_tile(color = "white") +
  geom_text(aes(label = Freq)) +
  scale_fill_gradient2(low = "red", high = "blue", mid = "white", 
                       midpoint = 0, limit = c(-1, 1), name = "Correlation") +
  theme_minimal() +
  theme(axis.text.x = element_text(angle = 45, vjust = 1, hjust = 1)) +
  labs(title = "Correlation Heatmap of Numeric Variables",
       x = "", y = "")

Error in placement_df %>% select(CGPA, SSC_Marks, HSC_Marks, AptitudeTestScore, : could not find function "%>%"
Traceback:

Interpretation: This correlation heatmap above helps us investigate any potential multicollinearity among the numeric explanatory variables. Most of the correlations lie in the moderate range of 0.4 to 0.5, indicating that they may have some association but not extreme collinearity. The highest observed correlation is between Aptitude Test Score and HSC Marks of 0.57, which is still within a tolerable range. This reassures us that including all of these variables in the model probably won't cause multicollinearity issues, although we could confirm this with VIF if nee

4. Method and Plan

Proposed Method:

I will use a logistic regression model to explore my research question. This model is ideal for the response variable PlacementStatus, a binary variable indicating whether or not a student got placed. The explanatory variables include academic performance (CGPA), skill-based attributes (SoftSkillsRating, AptitudeTestScore), and experience-related factors (Internships, Projects, PlacementTraining, and ExtracurricularActivities), etc.

Why this is an appropriate method:

Logistic regression is best suited for modelling the relationship between one or more explanatory variables and a binary response variable. It will allow us to assess how each explanatory factor is associated with the likelihood of a student being placed. We will do this by estimating odds ratios and evaluating statistical significance. This method helps us with our inference-based focus of the analysis as it will help us understand associations.

Assumptions required for this method:

• Each student case (observation) is independent of each other
• Log-odds of the response are linearly related to the explanatory variables
• All important variables are included in the dataset

Potential Limitations of this method:

• Logistic Regression only reveals association and does not reveal any causation
• May perform poorly if there is class imbalance
• Intercations and/or nonlinear effects will need to be explicitly modeled

5. Computational Code and Output

# Setting seed to make sure training and testing is reproducible
set.seed(2)

# Split data into train and test sets, using 70% training and 30% testing
train_df <- placement_df %>% 
    slice_sample(prop = 0.7)

test_df <- placement_df %>% 
    anti_join(train_df, by = "StudentID")

#Fitting logistic regression model to predict PlacementStatus, using all variables barring StudentID
# StudentID is excluded because it's just an identifier
full_model <- glm(PlacementStatus ~ . -StudentID, data = train_df, family = binomial)

# Tidy up the model output for easier interpretation
tidy_table <- tidy(full_model, conf.int = TRUE, exponentiate = TRUE) %>%
  filter(term != "(Intercept)") %>%
  mutate(
    p.value = case_when(
      p.value < 0.001 ~ "<0.001",
      TRUE ~ as.character(round(p.value, 3))
    ),
    estimate = round(estimate, 3),
    std.error = round(std.error, 3),
    statistic = round(statistic, 2),
    conf.low = round(conf.low, 3),
    conf.high = round(conf.high, 3)
  ) %>%
  arrange(desc(estimate))

tidy_table

# Created odds ratio plot to display the odds ratios

odds_ratio_plot <- ggplot(tidy_table, aes(x = estimate, y = term)) +
  geom_vline(xintercept = 1, linetype = "dashed", color = "red") +
  geom_point(size = 3) +
  geom_errorbarh(aes(xmin = conf.low, xmax = conf.high), height = 0.2) +
  scale_x_log10() + 
  labs(
    title = "Odds Ratios for Placement Status Prediction",
    subtitle = "1= No Effect, Values > 1 indicate higher odds of Placement",
    x = "Odds Ratio (log scale)",
  ) +
  theme_minimal() +
  theme(
    panel.grid.minor = element_blank(),
    axis.text.y = element_text(size = 10)
  )

odds_ratio_plot

# Predicting on the test set and graphing the ROC Curve

test_predictions <- predict(full_model, newdata = test_df, type = "response")
roc_curve <- roc(test_df$PlacementStatus, test_predictions)
plot(roc_curve, print.auc = TRUE, col = "blue", main = "ROC Curve")

A tibble: 10 × 7

term	estimate	std.error	statistic	p.value	conf.low	conf.high
<chr>	<dbl>	<dbl>	<dbl>	<chr>	<dbl>	<dbl>
PlacementTraining	2.460	0.085	10.59	<0.001	2.084	2.908
SoftSkillsRating	2.280	0.102	8.11	<0.001	1.869	2.784
ExtracurricularActivities	2.181	0.080	9.70	<0.001	1.863	2.554
CGPA	1.442	0.060	6.07	<0.001	1.282	1.624
Projects	1.287	0.045	5.65	<0.001	1.179	1.405
Workshops.Certifications	1.181	0.038	4.35	<0.001	1.096	1.273
AptitudeTestScore	1.072	0.006	12.52	<0.001	1.061	1.084
HSC_Marks	1.030	0.005	6.49	<0.001	1.021	1.039
SSC_Marks	1.026	0.004	6.68	<0.001	1.018	1.033
Internships	0.963	0.051	-0.74	0.459	0.872	1.064

Setting levels: control = 0, case = 1

Setting direction: controls < cases

Interpretation of the model and Issues:

In the full logistic regression model, PlacementTraining has the highest coefficient, indicating that it is the variable most associated with whether a student gets a placement or not. Internships is the only variable that doesn't seem to be statistically significant, which is unexpected given its assumed importance in real-world placement. This could be because very few students had more than one internship, which could have limited the model; converting Internships to a binary variable (Done at least one internship vs No internships) could help with this issue. Since all the other variables seem to be very significant, this could point to multicollinearity or overlapping effects, so we should experiment with interaction terms to help uncover some more meaningful relationships.

Based on the Odds Ratio plot, the variables further to the right have more of a positive effect on placement. Internships tend to have a negative effect on placement (Odds Ratio lower than 1), while SSC_Marks, HSC_Marks, and AptitudeTestScore appear to not have much influence (Odds Ratio around 1). Workshops.Certifications, Projects, and CGPA have a moderate positive impact on placement (Odds Ratio around 1.2-1.6). The variables with the strongest influence on placement are PlacementTraining, SoftSkillsRating, and ExtracurricularActivities, with an Odds Ratio of above 2, all having a large chance of increasing the odds of successful placement.

Based on the ROC curve and the AUC value of around 0.8777 on the test set, this indicates that the predictive performance of the model is very good.

6. Discussion

The modelling results point to several actionable lessons for universities and career‑services teams. First, placement‑training programmes deliver the single largest boost to employability, so expanding access and perhaps making completion a programme requirement could meaningfully raise placement rates. Almost as influential are communication and teamwork skills: embedding workshops and graded presentations throughout the curriculum would help students develop these attributes. Participation in clubs, sports, or societies is another strong signal of success; institutions might therefore recognise extracurricular engagement on transcripts or through micro‑credentials. Finally, internships warrant closer scrutiny—not in terms of how many a student completes, but in how relevant and substantive each placement is. Collecting richer metadata on duration and role relevance would allow future models to distinguish high‑quality internships from box‑ticking exercises.

From a methodological standpoint, the next step is to refine the treatment of the Internships variable by collapsing it into a simple indicator of whether a student has any internship experience, then testing interactions with other practical‑experience measures such as Projects. Elastic‑net regularisation could stabilise coefficients and guard against any lingering multicollinearity, while probability‑calibration techniques (Platt scaling or isotonic regression) would make the model’s predicted probabilities more actionable for counsellors who need clear cut‑off thresholds.

Looking ahead, richer data will enable deeper insights. Linking predictions to first‑year salary and job‑retention outcomes would reveal whether the drivers of placement also forecast longer‑term success. Incorporating interview scores or employer tier could help explain which student attributes convert applications into offers. Finally, a fairness audit—examining performance across gender, socioeconomic status, or other protected attributes—will ensure that model‑based recommendations do not inadvertently disadvantage any subgroup.

7. References

Cohen Freue, Gabriela. Course Materials. STAT 301 folder, Jupyter Notebooks, University of British Columbia, n.d.

Kumbhar, Ruchika. "Placement Prediction Dataset." Kaggle, https://www.kaggle.com/datasets/ruchikakumbhar/placement-prediction-dataset/data. Accessed 10 Apr. 2025.

The initial model and code up to the tidy table was taken from Vedant's Assignment 2.